Quartz 4

❯

ESM-IF

Created Apr 10, 2026Modified Apr 21, 2026

ESM-IF is an Inverse folding method that uses twelve million AlphaFold2 structures during training, in addition to PDB. It uses an encoder-decoder Transformer architecture with a GVP as well as fixed-order decoding (autoregressive).

Ref (1)

1.

Hsu C, Verkuil R, Liu J, Lin Z, Hie B, Sercu T, et al. Learning inverse folding from millions of predicted structures. openRxiv; 2022. Available from: https://doi.org/10.1101/2022.04.10.487779

Graph View

Backlinks

A subset of residues is sufficient to predict protein dG
AA identity can be determined from structural context alone
Adding noise while training non-Ab inverse folding models improves self-consistency while worsening sequence recovery
Antibody LMs outperform generic PLMs on intrafamily thermostability prediction
Averaging logits from multiple sources can improve fitness prediction
ESM-IF outperforms ProteinMPNN on Ab inverse folding and affinity prediction
ESM-IF representations can predict if residue binds ligands
ESM-IF, but not other inverse folding models, has learned some evolutionary constraints from sequence databases
Geometric Vector Perceptrons
Half of all disease-causing missense variants due to disruption of protein stability
Inverse folding NNs are better predictors of equilibrium dynamics than protein folding NNs
Most ML quality metrics cannot effectively predict enzyme activity after controlling for similarity to native
ProteinMPNN
Sequence- and structure-derived ML quality metrics from ML do not correlate with each other
Structure-based methods outperform sequence-based methods on protein stability prediction of point mutants, but not full sequences
Training inverse folding and diffusion models exclusively on predicted protein structures worsens performance due to how locally perfect they are
Inverse folding

GitHub
Discord Community

Quartz 4

Explorer

ESM-IF

Graph View

Backlinks